- Customer: Just Retirement Life (South Africa) Limited (https://justsa.co.za/)
- Industry: Financial Services (Life Insurance)
- Solution: Agentic Workflow Automation.
Challenge: High volumes of manual email requests for payslips, tax certificates, and increase letters created compounding operational strain — 8–12 minutes per request, processing errors, and heavy reliance on support staff, particularly during peak periods and overnight hours when no one was available.
Solution: Deployed InfoAgent, an AI-powered automation solution that intelligently processes inbound emails, validates requests against CRM data, and retrieves documents securely through integrated APIs — removing manual intervention from the standard request lifecycle entirely.
The support team historically faced significant operational bottlenecks during high-volume periods — Monday mornings, month-end reporting cycles, and peak tax season. Requests for client payslips and policy details required manual intervention, creating a repetitive, time-intensive, and manual workflow.
✅ Resource Intensive: Each ticket required 8–12 minutes of manual effort for verification and retrieval
✅ Operational Inefficiency: Manual processing involved repetitive verification and document retrieval across multiple systems, making the process time-consuming and dependent on staff availability.
✅ System Silos: No integration between the ticketing system, CRM, and document repository. Staff had to navigate all three manually for every request.
Bangalore Softsell developed InfoAgent, an AI-powered background service that automates the support queue end-to-end by integrating with Freshdesk, OpenAI, CRM, and a Document API. When a request arrives, InfoAgent polls Freshdesk every 30 minutes, parses the email using AI, validates the request against CRM, retrieves the document, and sends it back as an automated reply — all without human involvement for standard requests.
Request Interpretation
Key fields like Client ID, Policy Number, requested month & year are extracted from unstructured email content using GPT-4o
Schema Validation
Extracted data is validated using NJsonSchema to ensure it matches required internal patterns before any downstream calls are made
CRM Validation
The sender’s authority to access the requested client’s documents is verified against CRM before any document is retrieved
Document Retrieval
- The Document API is called to fetch the requested documents.
Automated Response
- The document is attached and sent back via Freshdesk, closing the ticket without manual intervention
To safeguard sensitive financial data, InfoAgent incorporates a multi-layered defence strategy:
Azure AI Content Safety
Screens all inputs for prompt injection and malicious content
Tamper-Evident Requests
SHA-256 hashes prevent replay attacks
Prompt Integrity
Verifies the system prompt template is unmodified at runtime
NJsonSchema Validation
Ensures AI output adheres to required data structures
CRM Validation
- Restricts document access to authorized personnel only
HTML Sanitization
- Prevents injection attacks in outgoing communications
Audit Logging
- Immutable, PII-masked logs with full event trail to Azure Application Insights.
Speed without accuracy creates new problems. Before any document is retrieved, InfoAgent applies to a structured validation layer using the CRM as the source of truth. Here are three examples that illustrate the system’s discipline.
Adviser Authentication & Active Status
- When a request is received from an adviser, InfoAgent validates the sender by matching their name and email address against CRM records. It also verifies that the adviser is currently active.
- Control: Prevents impersonation and blocks access from inactive or exited advisers, ensuring only authorized users can initiate requests.
Client–Adviser Relationship Verification
- Even authenticated advisers are restricted to accessing documents only for clients within their assigned portfolio. InfoAgent cross-checks the policy details in CRM to confirm that the requesting adviser is linked to the client.
- Control: Ensures strict access boundaries and prevents unauthorized cross-client data access.
Optimize Frameworks
- InfoAgent supports requests from policyholders directly. When the sender is not identified as an adviser, the system switches to a separate validation flow, verifying identity using policy details or ID numbers.
- Control: Enables secure self-service for legitimate customers while preventing unauthorized third-party access.
These validations are not just technical safeguards they reflect the compliance and governance requirements of operating in a regulated financial services environment.
Alongside the automation engine, a purpose-built dashboard was delivered to give the support team and management real-time insight into what the InfoAgent is doing at any moment. The dashboard provides four views:
Summary Metrics
Total requests processed, total successful document deliveries, average response time in seconds, and estimated AI processing cost. This gives management a view of throughput and spend.
Status Breakdown
Counts by outcome: sent successfully, missing fields, failed, or flagged for manual review. A spike in any category is immediately visible.
Workflow State Distribution
Provides a detailed breakdown of requests across all processing stages from Started to final response generation. This visibility helps quickly identify where failures occur and enables faster troubleshooting and resolution.
Granular Ticket View
Individual ticket records with full metadata, AI parsing output, and failure notes. Support staff use this to triage the minority of requests that require manual attention, with full context already available.
AI Accuracy & Low-Confidence Handling
- InfoAgent relies on GPT-4o to interpret unstructured email content, which introduces the risk of incorrect data extraction.
- Mitigation: Each AI output is evaluated with confidence scoring, and low-confidence extractions are flagged for human review instead of automated processing. Prompt templates are protected using SHA-256 integrity checks, and all inputs are screened through Azure Content Safety to prevent malicious or unsafe inputs.
Fallback When Automation Cannot Proceed
- Not all requests can be successfully automated due to missing data, validation failures, or system errors.
- Mitigation: In such cases, tickets remain open in Freshdesk and are routed for manual review. This ensures no request is lost and guarantees continuity of service even when automation cannot proceed.
Post-live, real-world usage provided valuable insights that helped further strengthen the system. These are some issues we ran into after the InfoAgent went live.
External Dependency & Infrastructure Challenges
- Occasional dependency or infrastructure disruptions caused processing interruptions and inconsistent request handling.
- Fix: Persisted AI parsing results and enabled the system to resume from the point of failure instead of restarting the entire workflow.
Duplicate AI Calls
- Failed tickets were being reprocessed from the beginning, triggering repeated AI parsing and increasing both cost and latency.
- Fix: Persisted AI parsing results and enabled the system to resume from the point of failure instead of restarting the entire workflow.
Client Email Validation
- Strict email matching rules led to rejection of legitimate requests when users contacted from alternate or shared email addresses.
- Fix: Implemented a more flexible identity validation approach that verifies user authenticity without requiring an exact email match.
Lack of Alerting
- System-level failures were logged internally but not surfaced to the support team, leading to delayed awareness and response.
- Fix: Introduced real-time alerting for polling, AI processing, and system failures with actionable context for quicker resolution.
Time Zone Mismatch
- Differences between UTC and local system time caused incorrect ticket age calculations, resulting in occasional reprocessing of resolved tickets.
- Fix: Standardized time zone handling across all components to ensure accurate ticket tracking and processing.
The headline numbers tell part of the story. The fuller picture is what the team can now do with the capacity that has been freed.
Resource Optimisation
Before InfoAgent, the support team spent a significant portion of every day on mechanical tasks: reading emails, looking up clients, retrieving documents. That work required care but not expertise. Staff have now transitioned to managing complex, high-value cases and exceptions that requires their judgement & skills.
Processing Efficiency
Average processing time of 8–12 minutes per ticket has been reduced to handling only problematic tickets, which require human judgement.
24/7 Service Continuity
The overnight backlog which previously greeted staff every morning and set the tone for the entire day is gone. InfoAgent processes requests around the clock, and the queue is clear before anyone arrives.
Scalability Without Proportional Headcount Growth
Volume spikes during month-end, tax season and peak periods no longer require surge staffing. InfoAgent absorbs the increase without friction.
InfoAgent represents a shift from manual, effort-driven operations to AI-powered intelligent execution. By combining large language models, automated validation, and secure system integration, the organisation has:
- Reduced operational friction across the full request lifecycle
- Improved accuracy and compliance through structured, automated validation
- Enabled scalable growth without proportional headcount increases
- Given the operations team visibility and control they did not have before
This transformation demonstrates how AI can move beyond experimentation to deliver real, measurable business value. More importantly, it shows how a well-governed, production-grade AI system can continuously improve through real-world feedback, becoming more resilient and valuable over time.
.NET 8 (C#), ReactJS, Azure OpenAI, Azure AI Content Safety, Azure Key Vault, SQL Server, Azure DevOps